Analyzing Large Receptive Field Convolutional Networks for Distant Speech Recognition

Abstract

Despite significant efforts over the last few years to build a robust automatic speech recognition (ASR) system for different acoustic settings, the performance of the current state-of-the-art technologies significantly degrades in noisy reverberant environments. Convolutional Neural Networks (CNNs) have been successfully used to achieve substantial improvements in many speech processing applications including distant speech recognition (DSR). However, standard CNN architectures were not efficient in capturing long-term speech dynamics, which are essential in the design of a robust DSR system. In the present study, we address this issue by investigating variants of large receptive field CNNs (LRF-CNNs) which include deeply recursive networks, dilated convolutional neural networks, and stacked hourglass networks. To compare the efficacy of the aforementioned architectures with the standard CNN for Wall Street Journal (WSJ) corpus, we use a hybrid DNN-HMM based speech recognition system. We extend the study to evaluate the system performances for distant speech simulated using realistic room impulse responses (RIRs). Our experiments show that with fixed number of parameters across all architectures, the large receptive field networks show consistent improvements over the standard CNNs for distant speech. Amongst the explored LRF-CNNs, stacked hourglass network has shown improvements with a 8.9% relative reduction in word error rate (WER) and 10.7% relative improvement in frame accuracy compared to the standard CNNs for distant simulated speech signals.Index Termsdeeply recursive network, dilated convolutional network, large receptive field network, speech recognition, stacked hourglass network.

Investigation of Machine Learning Approaches for Traumatic Brain Injury Classification via EEG Assessment in Mice

“…Furthermore, different methods have been proposed to train deeper yet more generalized models, namely L1 and L2 regularizations, batch normalization, dropout layers, and more data-side solutions like data augmentation [20,21]. Owing to the ability to efficiently pick up intricate features of data, CNNs have been widely deployed in many complex machine learning tasks from image processing to speech recognition [22][23][24] . Figure 2 demonstrates our feature extraction and CNN pipeline, with a relatively shallow architecture.…”

“…Classical EEG oscillations range from below 1 Hz to around 100 Hz which have been extensively studied for various medical applications. These classical frequency bands include delta (0.5-4 Hz), theta (4-8 Hz), alpha (8)(9)(10)(11)(12), beta (13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30), and gamma (30-100 Hz) [5], each of which exhibits unique underlying physiological mechanisms. EEG is used to evaluate several brain disorders such as epilepsy, Alzheimer's disease and lesions of the brain which reveals as seizure or unusually slow waves in EEG.…”

Frame-Based Overlapping Speech Detection Using Convolutional Neural Networks

“…Human brain is capable of focusing on a single talker in a multi-speaker environment, recognizing both the identity of the talker and also the content of the speech. However, the performance of speech analysis technologies such as speaker diarization, identification and Automatic Speech Recognition (ASR) is adversely affected in presence of co-channel speech [2,3,4]. In speaker diarization, the existence of overlapping speech in the training dataset leads to generating impure speaker models, which increases diarization error [5].…”

Domain Expansion in DNN-Based Acoustic Models for Robust Speech Recognition

“…Current state-of-the-art neural network-based ASR systems have advanced to nearly human performance in several eval-This project was supported in part by AFRL under contract FA8750-15-1-0205, and partially by the University of Texas at Dallas from the Distinguished University Chair in Telecommunications Engineering held by J. H. L. Hansen. uation settings [1,2]; however, these systems perform poorly for domains 1 that are not included in the original training data [3,4,5,6]. For example, if we train an ASR system using a U.S. English dataset, the performance of the system significantly degrades for other English accents (e.g., Australian, Indian, and Hispanic).…”

“…uation settings [1,2]; however, these systems perform poorly for domains 1 that are not included in the original training data [3,4,5,6]. For example, if we train an ASR system using a U.S. English dataset, the performance of the system significantly degrades for other English accents (e.g., Australian, Indian, and Hispanic).…”